pre-miRBench Dataset and Benchmarking Results

This record contains benchmark v3 of pre-miRBench, the version reported in the revised manuscript: the training, validation and four test partitions, the inputs of the pre-miRBench model and of six published precursor-miRNA predictors, the trained models, the predictions, metrics and tables of the benchmark and of the genome-scale evaluation on four held-out genomes (Drosophila melanogaster, D. ananassae, D. yakuba and Gallus gallus), the comparison with MirMachine 2 and ncOrtho, and the Docker images of the six published predictors. In v3, positive and negative windows are placed by the same rule: each 200 nt window is centred on the terminal loop of an RNALfold hairpin and moved by a seeded shift of at most 10 nt. Version v1 of this record holds the benchmark of the first submission. Code, the released model and instructions to download these files or rebuild them from the raw MirGeneDB and genome data: https://github.com/dimostzim/pre-miRBench. README.md in this record lists every file. Version v3.1 has the same data files as version v3 (doi:10.5281/zenodo.23933248). It corrects the section of README.md on rebuilding the dataset and updates release_metadata.json and SHA256SUMS.